Canada Spends Logo

Explore our Budget 2026 analysis

Grant: Discovery Grants Program - Individual

University of Toronto

$52,000.00

Learning on Manifolds for Robotic Manipulation

FY 2023-24

Summary
For many years, we have been promised capable robots that are able to perform complex tasks in human-centric environments, such as cleaning up a kitchen. And yet today, mobile robotic manipulators (i.e., mobile platforms equipped with one or more robot arms and end-effectors) are far from ubiquitous. So where are they? What is holding back progress? There are two important answers that stand out: first, it is difficult to reason over the high-dimensional sensing and action spaces in which these robots must operate and, second, even when reasoning is possible, it is fraught with uncertainty that, when misestimated, can lead to potentially catastrophic failures. Recent successes in the field of deep learning have shown a way forward, toward the development of truly robust robotic systems. Finding analytical solutions to problems in high-dimensional spaces is often intractable - deep learning provides a way to discover and leverage lower-dimensional 'latent' representations that encode the statistics of interest. Many latent spaces inherently possess a manifold structure and have an associated metric, allowing distances and directions to be computed in a principled way. The structure and metric can be utilized to more quickly converge to feasible solutions and to enforce constraints, for example. The goal of the proposed research program is to advance the state of the art in robotic manipulation, helping to bridge the gap and to make flexible mobile robots much more common. The focus on manipulation is precedented. Recent studies estimate that the global robotics market will grow at an annual rate of between 8% and 14% over the next 10 years. While autonomous vehicles will account for some of these gains, versatile manipulation platforms will be game-changing as a growth driver for the robotics industry and for society. Supporting this goal, the research program consists of three project themes: learning on manifolds, uncertainty learning on manifolds, and manifold and metric learning. Research activities will build upon key results from differential geometry, statistical estimation theory, and deep learning. Work within the first project will extend results from the applicant's group on the use of Riemannian manifolds for kinematics and motion planning to learning-based systems. Research within the second project will explore uncertainty estimation in deep neural networks and examine ways to build uncertainty models that respect the manifold structure of the underlying domain. Activities within the third project will focus on learning the manifold structure itself from noisy, multimodal sensor data. These topics have been studied in other fields, but for many robotics problems, the appropriate dimension and topology of the latent space are unknown. In summary, by identifying and respecting the underlying structure of the domain, the proposed research program will improve the robustness, flexibility, and versatility of robot manipulation.
Program
Discovery Grants Program - Individual
Location
Ontario, Canada
Recipient Type
Academia
Type
Grant
Source
NSERC's Awards Data

As published

Application id
RGPIN-2023-05036
Name nom
Kelly, Jonathan
Department département
Institute for Aerospace Studies
Organization id
31
Institution établissement
University of Toronto
Province en
Ontario
Province fr
Ontario
Country en
CANADA
Country fr
CANADA
Fiscal year exercice financier
2023
Competition year année de concours
2023
Award amount
52000
Program id
RGPIN
Program name en
Discovery Grants Program - Individual
Program name fr
Programme de subventions à la découverte - individuelles
Group en
DISCOVERY RESEARCH
Group fr
RECHERCHE AXÉE SUR LA DÉCOUVERTE
Committee code
1507
Committee name en
Computer Science
Committee name fr
Sciences informatiques
Area of application code
500
Area of application group en
Manufacturing processes and products
Area of application group fr
Procédés et produits industriels
Area of application en
Manufacturing processes and products
Area of application fr
Procédés et produits industriels
Research subject code
2603
Research subject group en
Robotics
Research subject group fr
Computer vision
Research subject en
Robotique
Research subject fr
Vision artificielle
Application title
Learning on Manifolds for Robotic Manipulation
Keyword
Deep Learning; Differential Geometry; Geometric Learning; Lie Groups; Machine Learning; Manifolds; Riemannian Geometry; Robotic Manipulation; Robot Motion Planning; Robot Perception
Application summary
For many years, we have been promised capable robots that are able to perform complex tasks in human-centric environments, such as cleaning up a kitchen. And yet today, mobile robotic manipulators (i.e., mobile platforms equipped with one or more robot arms and end-effectors) are far from ubiquitous. So where are they? What is holding back progress? There are two important answers that stand out: first, it is difficult to reason over the high-dimensional sensing and action spaces in which these robots must operate and, second, even when reasoning is possible, it is fraught with uncertainty that, when misestimated, can lead to potentially catastrophic failures. Recent successes in the field of deep learning have shown a way forward, toward the development of truly robust robotic systems. Finding analytical solutions to problems in high-dimensional spaces is often intractable - deep learning provides a way to discover and leverage lower-dimensional 'latent' representations that encode the statistics of interest. Many latent spaces inherently possess a manifold structure and have an associated metric, allowing distances and directions to be computed in a principled way. The structure and metric can be utilized to more quickly converge to feasible solutions and to enforce constraints, for example. The goal of the proposed research program is to advance the state of the art in robotic manipulation, helping to bridge the gap and to make flexible mobile robots much more common. The focus on manipulation is precedented. Recent studies estimate that the global robotics market will grow at an annual rate of between 8% and 14% over the next 10 years. While autonomous vehicles will account for some of these gains, versatile manipulation platforms will be game-changing as a growth driver for the robotics industry and for society. Supporting this goal, the research program consists of three project themes: learning on manifolds, uncertainty learning on manifolds, and manifold and metric learning. Research activities will build upon key results from differential geometry, statistical estimation theory, and deep learning. Work within the first project will extend results from the applicant's group on the use of Riemannian manifolds for kinematics and motion planning to learning-based systems. Research within the second project will explore uncertainty estimation in deep neural networks and examine ways to build uncertainty models that respect the manifold structure of the underlying domain. Activities within the third project will focus on learning the manifold structure itself from noisy, multimodal sensor data. These topics have been studied in other fields, but for many robotics problems, the appropriate dimension and topology of the latent space are unknown. In summary, by identifying and respecting the underlying structure of the domain, the proposed research program will improve the robustness, flexibility, and versatility of robot manipulation.